Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis

AI summary

Overview

Research area: Computer-aided synthesis planning (CASP), specifically single-step retrosynthesis (SSRS) with large language models, evaluated through a chemical-plausibility metric rather than conventional top-K accuracy.

Technical level: Advanced. The paper assumes familiarity with retrosynthetic analysis, SMILES notation, LLM fine-tuning (supervised fine-tuning and GRPO-based reinforcement learning), and cheminformatics concepts such as reaction centers and functional-group signatures.

Scope: The paper introduces a Top-K prompting and training paradigm, builds an ultra-large reaction dataset of about 45.6 million verified reactions, and trains a plausibility-aware LLM (C3LM) that is benchmarked against proprietary and open-weight foundation models and conventional SSRS models on the URSA-expert-2026 and USPTO-50K-test-mini benchmarks.

What This Paper Is About

Single-step retrosynthesis predicts which precursor reactants could form a given target molecule in one reaction step. The task is intrinsically one-to-many: a target can often be disconnected in several independent, chemically valid ways, yet most prior evaluation and training protocols ask for a single answer. This paper argues that single-answer benchmarking understates model capability and proposes Top-K prompting as both an inference and a training paradigm, combined with ChemCensor-based rewards, to generate diverse and chemically plausible reaction predictions.

Key Contributions

1. Top-K prompting mode for LLM training and prompting, presented as a best practice for increasing the diversity of generated reactions. Top-1 mode prompts with 15 templates and collects one answer each; Top-K mode appends the instruction "Give me 15 different answers" and collects up to 15 reactant sets per template from 3 random templates per target.

2. CREED-CCV-2+USPTO-XL, an ultra-large dataset of approximately 45.6 million verified reactions, which after merging and deduplication contains 3,680,906 unique products and 45,649,785 unique reactions.

3. A new C3LM (Chemistry Constraint-Consistent Language Model) trained on the novel dataset, initialized from the LFM2 2.6B checkpoint, which outperforms conventional SSRS models and other LLMs on the URSA-expert-2026 benchmark according to ChemCensor-aggregated metrics.

4. A reaction-uniqueness comparison of LLMs and conventional tools, showing that the two model families explore complementary reaction spaces and identifying the current frontier of the SSRS task.

Main Findings

• Top-K beats Top-1 broadly: Most benchmarked LLMs benefit from the Top-1 to Top-K transition, and model rankings change. Grok-4.1 produces the most plausible and diverse reactions in Top-1 mode, while Gemini 3.1 Pro improves dramatically in the Top-15 mode and becomes the best LLM baseline. Only GPT 5.2 performs worse in Top-K mode according to Av. PT-Top-10 values, while GPT 5.1 and Claude Sonnet 4.6 also show inferior performance in Top-K mode on other ChemCensor-based metrics.

• Top-K training also helps: Training C3LM-LFM2-CREED-CCV+USPTO in Top-K mode instead of Top-1 mode raised Max/@3/@5/@10 by +0.30/+0.62/+0.70/+0.60 respectively, with the largest improvement (more than 2.5-fold) on Av. PT-Top-10, pointing to increased diversity. This stage brought the SFT C3LM to the performance level of top-tier proprietary models Gemini 3.1 Pro and GPT 5.5.

• More data helps: Moving from CREED-CCV+USPTO to CREED-CCV-2+USPTO-XL (a greater than 6-fold increase in training set size) improved Max/@3/@5/@10 by +0.12/+0.13/+0.17/+0.29 respectively.

• Reward fine-tuning beats conventional models: The ChemCensor reward alone allowed C3LM to beat conventional SSRS models on the out-of-distribution URSA-expert-2026 benchmark on Av. PT-Top-K CC metrics. Adding the novelty reward produced dominance across all ChemCensor metrics on that benchmark. CC-RFT added +0.04/+0.07/+0.06/+0.02 to Max/@3/@5/@10, and the novelty reward contributed a further +0.08/+0.06/+0.08/+0.08.

• Best C3LM scores: C3LM-LFM2-RFT-CC-NR reached Max 2.16, Av. PT-Top-3 1.94, Av. PT-Top-5 1.73 and Av. PT-Top-10 1.37 on URSA-expert-2026, and 4.28 / 3.01 / 2.51 / 1.85 on USPTO-50K-test-mini. For comparison, MHNreact reached 2.05 / 1.84 / 1.62 / 1.28 and 4.86 / 3.30 / 2.68 / 1.90, and LocalRetro 2.11 / 1.85 / 1.59 / 1.22 and 4.84 / 3.31 / 2.67 / 1.81.

• Data leakage caveat on USPTO-50K-test-mini: 481 of 497 product molecules in USPTO-50K-test-mini have no more than 3 reactions in the public USPTO-full, making Av. PT-Top-10 the least leakage-vulnerable metric there. The near-5 Av. PT-Max CC values of conventional SSRS models (exact match to a synthetic precedent) suggest potential leakage of USPTO-50K-based benchmarks into these models, so that metric should be treated cautiously.

• Complementary chemical spaces: MHNreact dominates most LLMs in producing plausible reactions (ChemCensor score > 0) that the LLMs did not generate. The only models producing more unique reactions than MHNreact are C3LM-LFM2-CREED-CCV-2+USPTO-XL (+0.4 reactions on average) and C3LM-LFM2-RFT-CC-NR (+0.3 reactions on average). Every LLM nonetheless predicted some plausible reactions MHNreact did not.

• Frontier values: Pooling all 30 benchmarked models gives the detectable frontier of 2.25 / 2.15 / 2.04 / 1.81 on URSA-expert-2026 and 4.91 / 4.12 / 3.69 / 3.04 on USPTO-50K-test-mini (Max / @3 / @5 / @10). All 17 general-purpose LLMs together (2.19 / 2.07 / 1.93 / 1.65 on URSA) slightly exceed all 5 C3LMs together (2.23 / 2.04 / 1.88 / 1.57) on the Top-10 comparison in the paper's discussion, while the paper reports there is still room for improvement of +0.44 for Av. PT-Top-10 and +0.31 for Av. PT-Top-5.

Methodology in Plain English

The authors treat retrosynthesis as a question-answering problem: give a model the product's chemical structure as a text string (SMILES), wrapped in one of 15 natural-language prompt templates adapted from MolInstructions, and ask it for the reactants.

Two prompting setups are compared. In Top-1 mode, each of 15 templates is used once and one answer is collected per template, yielding 15 reactant sets per target. In Top-K mode, the prompt is modified with the suffix "Give me 15 different answers", and 3 random templates are used per target with up to 15 reactant sets collected per template; the three reaction sets are scored separately and averaged.

Training data came from two main sources. USPTO is a decontaminated version of USPTO-full with about 897K unique products and about 951K reactions, but mostly one reaction per product. CREED-CCV-2 and USPTO-XL expand coverage by running products through a Virtual Synthesis Engine that enumerates possible disconnections, then verifying them with ChemCensor (v.1.1.1); ChemCensor checks whether a reaction's reaction center and functional-group context have support in a reference library of verified transformations.

C3LM was initialized from the LFM2 2.6B checkpoint and trained with basic reasoning for 50,000 steps. Training added SMILES-specific tokens to the vocabulary and augmented SMILES with non-canonical random traversal to improve chemical generalization. A second stage used online reinforcement learning with Group Relative Policy Optimization (GRPO), combining six reward components: thinking-format validity (weight 0.1), molecular syntax (0.5), ChemCensor score (1.0), Top-K uniqueness (0.2), Top-K matching (0.1), and a novelty score (1.0) rewarding reactants outside the training list that still score positively under ChemCensor.

Evaluation used the URSA-expert-2026 benchmark (100 expert-validated novel target molecules built to be disjoint from public reaction datasets) and the cost-efficient USPTO-50K-test-mini set (497 targets), scored with ChemCensor v.1.1.1 against synthetic precedents from USPTO-full and, supplementally, USPTO-full combined with the commercial Pistachio set. Reactions from conventional models were generated via Syntheseus.

Why This Matters

Research impact: The paper reframes how SSRS models should be benchmarked, arguing that single-answer protocols obscure the one-to-many nature of retrosynthesis. It also quantifies a frontier across 30 models, provides a plausible upper bound for current metrics, and shows that LLM and conventional model outputs are chemically complementary rather than redundant — a direct argument for ensemble systems. The ChemCensor source code is publicly available, and updated benchmarking results are published at a dedicated website.

Real-world applications:

• Medicinal chemistry: judging whether newly designed compounds are synthetically accessible before committing laboratory resources.
• Route planning in drug discovery: supplying a single-step disconnection model that a multistep search engine orchestrates into complete synthetic routes.
• Retrosynthesis benchmarking and model selection: providing a plausibility-based evaluation protocol for teams choosing between LLM and conventional tools.
• Ensemble synthesis planning: combining conventional SSRS models and LLMs to cover a broader space of plausible reactions.

Industry relevance: The work comes from Insilico Medicine and positions plausibility-aware LLM training as a practical direction for CASP systems used in pharmaceutical R&D. The paper explicitly acknowledges partial circularity between optimizing on ChemCensor-based rewards and evaluating on ChemCensor metrics, framing this as a deliberate design choice aimed at maximizing utility for practitioners, and it addresses dual-use and biosafety considerations for synthesis-planning tools.

Future Directions

1. Closing the remaining frontier gap: Pooled results from all 30 models leave headroom of +0.44 on Av. PT-Top-10 and +0.31 on Av. PT-Top-5, giving a quantitative target for the next model iteration.

2. Ensemble systems: Because LLM and conventional models explore complementary reaction spaces, the authors motivate ensembles of a conventional model and an LLM to cover a wider set of plausible reactions.

3. Beyond the plausibility proxy: ChemCensor does not account for practical laboratory parameters such as reaction conditions, solvents and purification methods, and the reference precedents are limited to patent-derived reaction space; extending plausibility modeling is left open.

4. Richer diversity metrics and broader chemical coverage: Chemical diversity is currently assessed only by exact SMILES string matching, without chemically grounded comparison such as reaction class or mechanistic type, and the template-based training data generation may be biased toward particular chemical patterns and may miss novel-chemistry transformations.

Target Audience

This paper is most useful for cheminformatics and machine learning researchers working on retrosynthesis and computer-aided synthesis planning, LLM practitioners interested in reinforcement learning with domain-specific verifiable rewards, and benchmark designers who need plausibility-based rather than accuracy-based evaluation protocols. Medicinal chemists and computational drug-discovery teams evaluating synthesis-planning tools will also find the practical framing relevant, though the technical level is advanced.

Authors’ abstract

Single-step retrosynthesis is a central component of computer-aided synthesis planning, yet its intrinsically one-to-many nature is poorly captured by single-answer evaluation and benchmarking protocols. To address this, we introduce Top-K prompting as a robust training and inference paradigm to better capture diverse, plausible reaction predictions. We compile CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of ~45.6 million verified reactions to train the C3LM (Chemistry Constraint-Consistent Language Model). By integrating fine-tuning with ChemCensor-based and novelty-oriented rewards, our model achieves state-of-the-art performance on the OOD URSA-expert-2026 benchmark. Further analysis of reaction uniqueness shows that LLMs and conventional models explore complementary reaction spaces, motivating ensemble-based retrosynthesis systems. Overall, our results establish Top-K, plausibility-aware training as a practical new direction for robust future LLM-based synthesis planning.